Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
Research Division Federal Reserve Bank of St. Louis Working Paper Series
Identifying Technology Shocks in the Frequency Domain
Riccardo DiCecio and
Michael T. Owyang
Working Paper 2010-025A http://research.stlouisfed.org/wp/2010/2010-025.pdf
August 2010
FEDERAL RESERVE BANK OF ST. LOUIS Research Division
P.O. Box 442 St. Louis, MO 63166
______________________________________________________________________________________
The views expressed are those of the individual authors and do not necessarily reflect official positions of the Federal Reserve Bank of St. Louis, the Federal Reserve System, or the Board of Governors.
Federal Reserve Bank of St. Louis Working Papers are preliminary materials circulated to stimulate discussion and critical comment. References in publications to Federal Reserve Bank of St. Louis Working Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors.
Identifying Technology Shocks
in the Frequency Domain1
Riccardo DiCecio
Federal Reserve Bank of St. Louis
Michael T. Owyang
Federal Reserve Bank of St. Louis
First Draft: June 2010
This Draft: August 5, 2010
1We are grateful to N. Francis, T. Lubik, V. Ramey, and H. Uhlig for comments and discussions.
Kristie M. Engemann provided research assistance. All errors are our own. Any views expressed
are our own and do not necessarily re�ect the views of the Federal Reserve Bank of St. Louis or
the Federal Reserve System. Corresponding author: Riccardo DiCecio, [email protected].
Abstract
Since Galà [1999], long-run restricted VARs have become the standard for identifying the
e¤ects of technology shocks. In a recent paper, Francis et al. [2008] proposed an alterna-
tive to identify technology as the shock that maximizes the forecast-error variance share of
labor productivity at long horizons. In this paper, we propose a variant of the Max Share
identi�cation, which focuses on maximizing the variance share of labor productivity in the
frequency domain. We consider the responses to technology shocks identi�ed from various
frequency bands. Two distinct technology shocks emerge. An expansionary shock increases
productivity, output, and hours at business-cycle frequencies. The technology shock that
maximizes productivity in the medium and long runs instead has clear contractionary e¤ects
on hours, while increasing output and productivity.
JEL: C32, C50, E32.
Keywords: Aggregate �uctuations; Business cycle; Frequency domain; Technology shocks;
Vector autoregressions.
1 Introduction
Estimated impulse responses identi�ed from VARs are often used to distinguish between
competing macroeconomic theories. For example, using long-run restrictions, Galà [1999]
and Francis and Ramey [2005] found that hours decline in response to a positive technology
shock and argued that this provided evidence invalidating the standard RBC model in favor
of sticky price models. This conclusion, however, appears to depend on the speci�cation
of hours in the VAR.1 The salient result is that if speci�ed in di¤erences, hours contract.
On the other hand, Christiano et al. [2003] found that if speci�ed in levels, hours respond
positively on impact. Arguing that hours per capita should be stationary fails to provide
any resolution: Pesavento and Rossi [2005] showed that when the hours process is presumed
stationary but close to a unit root, hours contract.
More recently, a few papers have attempted to resolve these con�icting results by exam-
ining the manner in which productivity and hours are detrended. Fernald [2007] proposed
that productivity growth and hours had simultaneous shifts in mean. He argued that the
shifts in productivity growth did not re�ect technological innovation and should, therefore,
be removed. One solution was to demean the series, while allowing for multiple breaks in
the intercept, which con�rmed the �nding of a contraction in hours.2 Francis and Ramey
[2009] argued that the apparent unit root in hours was caused by changing demographic
trends; once properly detrended, hours declined on impact.3 Gospodinov et al. [2009], on
the other hand, argued that if one is interested in preserving the underlying comovements
between productivity and hours, the VAR should be estimated with hours in levels and not
detrended. In this case, they found that hours rise following a technology shock. Thus,
1Others have used sign restrictions to identify technology shocks. Dedola and Neri [2007] employed
restrictions recovered from a DSGE model. Peersman and Straub [2007] used restrictions recovered from a
New Keynesian model. Both found neutral technology shocks to be expansionary on impact.2This is similar to the recommendation of Canova et al. [2010], who argue against detrending the hours
series.3Galà and Rabanal [2005] also showed that neutral technology shocks are contractionary when hours is
detrended.
1
how the VAR distinguishes between models appears to depend on whether one believes the
low-frequency comovements can be attributed to technology.
Although long-run restrictions have been the popular choice for identi�cation, macro
models also have implications for the e¤ect of technology shocks at non-zero frequencies.
For example, the RBC literature [see Prescott, 1986] typically assumes that the business-
cyle frequencies of the data (hours, output, investment, consumption) are driven solely by
technology shocks.4 Thus, an alternative to identifying technology shocks using long-run
restrictions is to impose restrictions on the e¤ect of technology shocks at non-zero frequencies.
In particular, we identify the shock that maximizes the share of the forecast-error variance
(FEV) in productivity growth and hours at various frequencies in a standard bivariate VAR.
We consider business-cycle frequencies identi�ed using linear �lters: the HP1600 �lter and
the band-pass (BP) �lter for periods between 8 and 32 quarters. We also consider medium-
term cycles identi�ed by the BP �lter for periods of either 32-80 or 80-200 quarters. Our
identi�cation is adapted to the frequency domain from the method introduced by Faust
[1998] and extended to technology shocks in Francis et al. [2008] (FOR).5 We analyze the
e¤ect of varying the frequency window on the identi�ed the shocks. As in Fernald [2007],
we account for the low-frequency comovement between productivity growth and hours by
demeaning the two series and allowing for two breaks in mean.
Two distinct technology shocks emerge. When identi�ed from frequencies lower than
the business cycle, technology shocks have Schumpeterian e¤ects [Caballero and Hammour,
1994, 1996, Canova et al., 2007]. While productivity increases persistently, output responds
ambiguously on impact, with both positive and negative draws, and increases eventually.
These shocks have a clear contractionary e¤ect on hours, which fall on impact and converge to
4Hansen [1997] pointed out that RBC models driven by persistent, but stationary, technology shocks and
models driven by I(1) shocks have similar implications at business-cycle frequencies.5The identi�cation in FOR is similar in �avor to Uhlig�s [2004] medium-run identi�cation. Uhlig argued
that other shocks (e.g., dividend tax shocks) could have long-run e¤ects on productivity and, thus, confound
the Galà [1999] identi�cation. Like FOR, Uhlig computed the FEV share but, instead, calibrated it to a
value determined by Monte Carlo simulations from a theoretical model.
2
zero from below. Interestingly, the shocks that drive productivity in the medium term (32-80
and 80-200 quarters) are indistinguishable from each other and from the shock identi�ed with
long-run restrictions. On the other hand, when identi�ed from business-cycle frequencies,
technology shocks are expansionary, inducing positive comovement of ouptut, hours, and
productivity, and are broadly consistent with the predictions of an RBC model. The shocks
identi�ed by the HP1600 and the BP8;32 �lters are almost identical.
The balance of the paper is organized as follows: Section 2 describes the baseline struc-
tural VAR and the identi�cation problem. In this section, we also discuss the issues of
inference using long-run restrictions in small samples. Section 3 presents our alternative
frequency-based identi�cation. Section 4 compares the results across di¤erent frequency
windows used for identi�cation. Section 5 summarizes and o¤ers some conclusions.
2 The Baseline Structural VAR
Consider the reduced-form VAR(p) representation of the (n� 1) vector of variables Yt:
Yt = A (L)Yt�1 + ut; (1)
where A (L) is a matrix polynomial in the lag operator, L, of order p. The reduced-form
residuals, ut, have zero mean, i.e., E (ut) = 0, and covariance matrix �, i.e., E (utu0t) = �.
We can rewrite (1) as
Yt = A (L)Yt�1 + C"t;
where C � (A0)�1. The structural matrix, A0, maps the reduced-form residuals into the
structural shocks, "t, where E ("t"0t) = I.
The central issue in the structural VAR literature involves the identi�cation of the struc-
tural matrix, A0. It is well known that in�nitely many possible structural matrices can be
derived from the decomposition of the reduced-form variance-covariance matrix, �. Various
methods have been used to impose su¢ cient identifying restrictions. Short-run identi�ca-
3
tions, for example, impose restrictions directly on the elements of the structural matrix.
While short-run identi�cation schemes have proven less controversial econometrically, they
typically are imposed through (potentially ad hoc) recursive ordering restrictions and may,
therefore, be less economically appealing.
Long-run restrictions, on the other hand, are often more consistent with theoretical mod-
els. For example, technology shocks are often thought of as important drivers of business
cycles in RBC models. Thus, the empirical literature has used VARs to identify technology
shocks assuming that they are the sole source of the unit root in labor productivity [see,
e.g., GalÃ, 1999]. This identifying assumption is implemented in VARs by imposing that non-
technology shocks have no long-run� i.e., the frequency-zero� e¤ect on labor productivity.
Long-run restrictions identify the structural matrix, A0, using restrictions of the form
�[I � A (1)]�1C
ij= 0; (2)
which impose that shock j has no long-run e¤ect on variable i.
Long-run restrictions� despite their theoretical appeal� have been shown to have some
drawbacks. Based on arguments initially made by Sims [1972] and Faust and Leeper [1997],
Christiano et al. [2006] contended that, in general, biases in the impulse responses of a
VAR depend on two components: propagation (i.e., misspecifying the companion matrix
for the VAR) and identi�cation of the structural matrix (i.e., incorrectly choosing C). The
former should be invariant to the identi�cation strategy. Christiano et al. [2006] advocated
for short-run restrictions on the basis that identi�cation of C relies on the estimate of �
only. Identi�cation of C via long-run restrictions, on the other hand, relies on the estimate
of � and the estimates of the As via the use of A (1) in (2). Because the VAR is truncated,
the estimate of A (1) may have greater bias when the actual share of the spectral density
of the data is not large near frequency zero. That is, we may encounter inference problems
when we use the spectral density at the zero frequency as an input to the identi�cation if
there is not much information at the zero frequency.
In many models, there is a long-run interpretation of the shocks, and the spectral density
4
near frequency zero may be su¢ ciently high to accurately estimate A (1). But what happens
when it is not? The solution proposed by Christiano et al. [2006] is an alternative estimator,
based on the Bartlett estimator which is aimed at estimating A (1) more accurately. Another
alternative advocated by FOR is to examine theoretically-consistent identifying restrictions
that do not depend so heavily on the zero frequency.
3 Identifying Shocks in the Frequency Domain
We describe how to identify the structural shock that maximizes (or minimizes) the share
of the FEV of a given variable in the frequency domain.6 The method can be applied to
any frequency window, including the entire spectrum. De�ne the matrix D as the Cholesky
factor of � and & i as the ith colum vector of D. We will refer to the set of & i as the Cholesky
shocks. Then, all of the possible impulse vectors, � (�), resulting from exact identi�cation
can be expressed as a linear combination of the Cholesky shocks as follows: � (�) = �0D,
where (�0�) = 1. Put di¤erently, � is a column of a rotation matrix that maps the Cholesky
shocks into another set of structural shocks. The goal of the procedure that follows is to
generate �.
In particular, we consider the component of the FEV of the variable of interest identi�ed
by a linear �lter, � (!). For convenience, we can �rst map the variables in Yt into other
variables of interest in the vectors ~Yt and �Yt as follows:
�Yt = F (L)�1 Yt; (3)
~Yt = G (L) �Yt = G (L)F (L)�1 Yt; (4)
where F (L) and G (L) are matrix polynomials in the lag operator. The matrix F (L) may
be used, for example, to convert some of the elements of Yt into di¤erences, while G (L)
rede�nes some variables (e.g., transforming output and hours into labor productivity).
6The construction of the FEV shares in the frequency domain follows Altig et al. [2005].
5
The spectral density of ~Yt is given by:
S ~Y�e�i!
�= H
�e�i!
��H
�e�i!
�0; (5)
where H (e�i!) �nG (e�i!)F (e�i!)
�1[I � A (e�i!)]�1
o. Given a structural shock (i.e., a
vector �) and a linear �lter with gain � (!), the share of the FEV of variable j due to � is
given by7:
V ~Yj (�) = �0
hR ��� � (!)S �Yj (e
�i!) d!i
trhR ��� � (!)S �Yj (e
�i!) d!i�: (6)
The shock with the largest share of the FEV of variable j associated with the �lter � (!)
is given by �� = argmax� V ~Yj (�). In principle, the choice of the �lter, � (!), can be ad
hoc (e.g., HP �lter) or governed by theory. For example, the standard long-run restriction
would amount to maximizing the FEV as ! approaches zero. In practice, we can utilize a
set of linear �lters constructed such that the gain re�ects, say, business-cycle or medium-run
frequencies.
4 Technology Shocks
In this section, we analyze the e¤ects of technology shocks in a bivariate VAR in labor
productivity growth,� log (Y=H), and hours per capita, log (H=N).8 We consider technology
shocks identi�ed by the algorithm proposed in Section 3 for a variety of frequency windows.
The sample covers the period 1948:Q2-2009:Q4. Previous studies have highlighted changes
in the mean growth of productivity over the sample. We follow Fernald [2007] and demean
7The integrals in equation (6) can be computed by Gauss-Legendre quadrature [see Press et al., 2007, p.
183].8Productivity growth is measured as the annualized log-di¤erence of output per hour of all persons in the
nonfarm business sector; hours per capita are included in logs and measured as hours of all persons in the
nonfarm business sector scaled by the civilian noninstitutional population (16 years and over). The Haver
mnemonics in the USECON database for productivity, hours, and population are LXNFA, LXNFH, and
LN16N, respectively.
6
both productivity and hours, allowing for two breaks: 1973:Q2 and 1997:Q2 (see Table 1).
Figure 1 portrays the two time series and their corresponding sample spectral densities after
demeaning.
Using the notation introduced above, we have:
Yt =
�� log
�Y
H
�; log
�H
N
��0;
�Yt =
�log
�Y
H
�; log
�H
N
��0= F (L)�1 Yt;
~Yt =
�log
�Y
N
�; log
�Y
H
�; log
�H
N
��0= G (L) �Yt:
The matrix polynomials, F (L) and G (L), that allow to recover productivity, log (Y=H), and
output per capita, log (Y=N), are given by:
F (L) =
24 1� L 0
0 1
35 ;
G (L) =
266641 �1
1 0
0 1
37775 :The reduced-form VAR is estimated with four lags, i.e., p = 4.
To estimate the reduced form of the VAR, we utilize Bayesian techniques. For the Gibbs
sampler, we impose a Je¤reys [1961] prior on the reduced-form VAR parameters9:
p (A;�) / det (�)�n+12 ; (7)
where A = [A1; :::; Ap]0. The posterior distribution of the reduced-form VAR parameters
belongs to the inverse Wishart-Normal family:
(�jYt=1;:::;T ) � IW�T �; T � k
�; (8)
(Aj�; Yt=1;::;T ) � N�A;� (X 0X)
�1�; (9)
9We truncate the prior to rule out posterior draws that imply explosive roots in the VAR.
7
where A and � are the OLS estimates of A and �, T is the sample length, k = (np+ 1),
and X is de�ned as
X =hx01; :::; x
0
T
i0;
x0t =h1; Y 0t�1; :::; Y
0
t�p
i0:
We identify �ve di¤erent �Max Share� technology shocks that maximize the FEV of
productivity at di¤erent frequencies:
1. the business-cycle frequencies extracted by an HP �lter of parameter � = 1; 600;
2. the business-cycle frequencies with a period between 8 and 32 quarters;
3. the frequencies corresponding to �uctuations with a �medium-run�period between 32
and 80 quarters;
4. the frequencies corresponding to �uctuations with a �medium-run�period between 80
and 200 quarters; and
5. the entire set of frequencies in [��; �] :
The gain functions of the corresponding linear �lters are reported in Table 2. The fre-
quencies corresponding to a given interval of periods, [l; u], are extracted with a BP �lter
[Christiano and Fitzgerald, 2003]. For comparison, we also consider the long-run identi�ca-
tion strategy proposed by Galà [1999].
For each draw from the posterior distribution of the reduced-form VAR parameters given
by equations (8) and (9), we identify a technology shock and construct impulse response
functions (IRFs) and FEV shares. For each of the identi�cation schemes described above,
we obtain 5; 000 draws from the posterior distribution of IRFs to various technology shocks
and the corresponding FEV shares at various horizons.
8
4.1 Results
We plot the median, 16th, and 84th percentiles of the posterior distributions of the impulse
responses to a one-percent increase in productivity growth in Figures 2-3. As a basis for
comparison, the impulse responses to the technology shock identi�ed by the standard long-
run restrictions are shown in the last column of Figure 3.
Two distinct technology shocks emerge from our identi�cation schemes. The shock that
maximizes the FEV of productivity at business-cycle frequencies� BP8;32 and HP1600 produce
almost indistinguishable IRFs� resemble the shock that drives RBC models. It induces
positive comovement between productivity, output, and hours. The responses of output and
hours are persistent and hump-shaped.
When maximizing the FEV of productivity at lower frequencies, the same shock emerges
from di¤erent identi�cation schemes� BP32;80, BP80;200, the entire set of frequencies, and
long-run restrictions. On impact this shock generates an ambiguous output response with
both positive and negative draws; eventually, output increases. This shock has a clear
contractionary e¤ect on hours: The median IRF is negative on impact and converges to zero
from below. The persistent increase in productivity is similar to the one generated by the
business-cycle shock described above.
Recall that the identi�ed technology shock is a linear combination of the Cholesky shocks,
where ��� the appropriate column of a rotation matrix� gives the weights. Figure 4 shows
the responses to the two Cholesky shocks. Figure 5 shows the posterior distributions of the
rotation angle between the two Cholesky shocks, �� = atan2 (��2; ��1), where �
�1 and �
�2 are
the two elements of ��. First, note that, for the bivariate system, the �rst Cholesky shock
(shown in the �rst column of Figure 4) appears consistent with the shock taken from the
standard RBC model. This would suggest that, for the shock identi�ed by the HP1600 or
the BP8;32 �lter, the rotation angle between the two Cholesky shocks should be centered at
zero. The �rst panel of Figure 5 reveals this to be true. For all the other frequency windows
we consider, more weight is placed on (the negative of) the second Cholesky shock. The
negative of this shock is similar to a contractionary technology shock.
9
Tables 3-5 show how much of the FEV of each variable the shocks explain at various
horizons. The last row of Table 3 reports the Max Share of productivty associated with each
shock, Vlog(Y=H) (��). Both the business-cycle and low-frequency technology shocks explain
most of the FEV of productivity at all horizons. The low-frequency shock�s share approaches
100% as the horizon increases. In contrast, the business-cycle shock explains almost all of
productivity at short horizons and more than 60% at long horizons.
The business-cycle frequency shock explains more than 50% of the FEV of output at all
horizons. The medium-term/low-frequency shock explains very little of the FEV of output
at short horizons� less than 15% up to two years. The FEV share of this shock increases
above 70% at long horizons.
Both shocks we identify account for less than 25% of the FEV of hours at any horizon.
The �nding that technology shocks account for little of the variance of hours is common-
place in the literature. The FEV share of the business-cycle (low-frequency) shock increases
(decreases) with the forecast horizon.
5 Conclusions
Since Galà [1999], technology shocks have been identi�ed in VARs using long-run restrictions.
In a recent paper, FOR� building on Faust [1998]� proposed a �nite-horizon alternative to
the long-run identi�cation. Here, we extend their �nite-horizon alternative to the frequency
domain. We identify technology shocks by the rotation of the Cholesky factor that maximizes
the FEV for a chosen frequency window. The frequency domain provides natural identifying
restrictions as many macro models have implications for technology shocks at business-cycle
frequencies.
Using a bivariate VAR with productivity and hours, we compare the identi�cation over
a number of frequency windows. The resulting impulse responses are broadly consistent
for frequency windows outside the business cycle. The shock identi�ed by these frequencies
is contractionary in hours with essentially zero e¤ect on output on impact. On the other
10
hand, the shock identi�ed by maximizing the FEV for business-cycle frequencies is more
consistent with an RBC technology shock� hours rise weakly on impact and output rises
unambiguously.
Another fundamental di¤erence between the two types of shocks identi�ed by the di¤erent
frequency windows is at what horizon the shock has the greatest e¤ect. For the contrac-
tionary shock, technology accounts for a relatively small share of the short-horizon FEV of
productivity and output. At longer horizons, this shock explains more of the variation in
both those variables. While the shock never explains a large portion of the FEV in hours, the
amount explained declines as the forecast horizon grows. The shock identi�ed by business-
cycle frequencies� the shock most consistent with the RBC shock� behaves in the opposite
manner, explaining more of productivity and output at short horizons and accounting for
less variation as the horizon grows.
11
References
D. Altig, L. J. Christiano, M. Eichenbaum, and J. Linde. Technical appendix to: Firm-
speci�c capital, nominal rigidities and the business cycle. URL http://faculty.
wcas.northwestern.edu/~lchrist/research/ACEL/technical.pdf. unpublished man-
uscript, Northwestern University, July 2005.
R. J. Caballero and M. L. Hammour. The cleansing e¤ect of recessions. American Economic
Review, 84(5):1350�68, December 1994.
R. J. Caballero and M. L. Hammour. On the timing and e¢ ciency of creative destruction.
Quarterly Journal of Economics, 111(3):805�52, August 1996.
F. Canova, D. Lopez-Salido, and C. Michelacci. Schumpterian technology shocks. unpub-
lished manuscript, CREi, Barcelona, Spain, November 2007.
F. Canova, D. Lopez-Salido, and C. Michelacci. The e¤ects of technology shocks on hours
and output: a robustness analysis. Journal of Applied Econometrics, 25(5):755�73, August
2010.
L. J. Christiano and T. J. Fitzgerald. The band pass �lter. International Economic Review,
44(2):435�465, May 2003.
L. J. Christiano, M. Eichenbaum, and R. Vigfusson. What happens after a technology shock?
NBER Working Paper No. 9819, July 2003.
L. J. Christiano, M. Eichenbaum, and R. Vigfusson. Assessing structural VARs. NBER
Macroeconomics Annual, 21:1�71, 2006.
L. Dedola and S. Neri. What does a technology shock do? a var analysis with model-based
sign restrictions. Journal of Monetary Economics, 54(2):512�49, March 2007.
J. Faust. The robustness of identi�ed VAR conclusions about money. Carnegie-Rochester
Conference Series on Public Policy, 49:207�44, 1998.
12
J. Faust and E. M. Leeper. When do long-run identifying restrictions give reliable results?
Journal of Business and Economic Statistics, 15(3):345�53, 1997.
J. G. Fernald. Trend breaks, long-run restrictions, and contractionary technology improve-
ments. Journal of Monetary Economics, 54(8):2467�85, November 2007.
N. Francis and V. A. Ramey. Is the technology-driven real business cycle hypothesis dead?
shocks and aggregate �uctuations revisited. Journal of Monetary Economics, 52(8):1379�
99, November 2005.
N. Francis and V. A. Ramey. Measures of per capita hours and their implications for the
technology-hours debate. Journal of Money Credit and Banking, 41(6):1071�97, September
2009.
N. Francis, M. T. Owyang, and J. E. Roush. A �exible, �nite-horizon identi�cation of
technology shocks. Federal Reserve Bank of St. Louis Working Paper No. 2005-024E, July
2008.
J. GalÃ. Technology, employment, and the business cycle: Do technology shocks explain
aggregate �uctuations? American Economic Review, 89(1):249�71, March 1999.
J. Galà and P. Rabanal. Technology shocks and aggregate �uctuations: How well does
the RBC model �t postwar U.S. data? In M. Gertler and K. Rogo¤, editors, NBER
Macroeconomics Annual 2004, pages 225�88, Cambridge, MA, 2005. MIT Press.
N. Gospodinov, A. Maynard, and E. Pesavento. Sensitivity of impulse responses to small
low frequency co-movements: Reconciling the evidence on the e¤ects of technology shocks.
CIREQ Cahier 03-2009, March 2009.
G. D. Hansen. Technical progress and aggregate �uctuations. Journal of Economic Dynamics
and Control, 21(6):1005�23, June 1997.
H. Je¤reys. Theory of Probability. Oxford University Press, London, 3rd edition, 1961.
13
G. Peersman and R. Straub. Putting the new Keynesian model to a test. unpublished
manuscript, Ghent University, June 2007.
E. Pesavento and B. Rossi. Do technology shocks drive hours up or down? a little evidence
from an agnostic procedure. Macroeconomic Dynamics, 9(4):478�88, September 2005.
E. C. Prescott. Theory ahead of business cycle measurement. Federal Reserve Bank of
Minneapolis Quarterly Review, 10(4):9�22, Fall 1986.
W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes.
Cambridge University Press, New York, third edition, 2007.
C. Sims. Money, income, and causality. American Economic Review, 62(4):540�52, Septem-
ber 1972.
H. Uhlig. Do technology shocks lead to a fall in total hours worked? Journal of the European
Economic Association, 2(2-3):361�71, April/May 2004.
14
Sampe period
Variable1948:Q2-
1973:Q1
1973:Q2-
1997:Q1
1997:Q2-
2009:Q4
400�� log�YH
�2:78 1:33 2:94
log�HN
��7:54 �7:56 �7:53
Table 1: Average productivity growth and hours p.c.
Filter � (!)
HP�4�[1�cos(!)]2
1+4�[1�cos(!)]2
BPl;u 1[ 2�u ;2�l ]
All 1[��;�]
Table 2: Gains of various �lters
Max Share Long
Horizon HP1600 BP8;32 BP32;80 BP80;200 All run
1 99:49[97:60;99:95]
98:04[90:85;99:03]
63:91[48:92;76:97]
62:20[37:13;83:30]
63:25[27:45;90:84]
60:94[19:75;92:53]
4 96:78[94:65;98:12]
96:15[91:02;97:87]
71:19[58:71;81:11]
70:00[46:76;86:79]
70:19[36:80;92:12]
68:10[27:98;92:88]
8 82:09[72:25;89:74]
83:91[70:67;91:09]
83:22[76:35;88:70]
82:13[68:25;90:08]
80:28[58:71;89:71]
79:01[51:69;89:51]
20 71:40[50:10;88:06]
73:60[46:42;91:46]
90:56[86:37;93:47]
91:03[85:36;94:82]
89:28[78:56;93:81]
88:50[73:92;93:52]
40 67:97[36:66;91:02]
69:39[32:93;93:33]
92:89[85:25;95:72]
94:83[91:61;96:82]
94:20[89:13;96:68]
93:82[86:46;96:50]
80 65:73[26:73;92:65]
66:32[22:58;94:07]
94:07[79:63;97:28]
96:64[93:38;97:96]
96:94[94:39;98:27]
96:85[93:63;98:21]
200 63:56[20:61;93:83]
63:38[16:39;94:91]
94:28[74:09;98:53]
97:99[92:99;98:97]
98:66[97:16;99:26]
98:74[97:65;99:28]
Vlog(Y=H) (��) 77:27
[70:24;83:82]72:29
[64:35;80:57]93:63
[83:89;97:93]97:85
[90:88;99:73]98:70
[97:76;99:18]100[�]
Table 3: FEV share of productivity due to technology shocks.
15
Max Share Long
Horizon HP1600 BP8;32 BP32;80 BP80;200 All run
1 59:94[48:62;70:32]
57:08[36:96;76:67]
6:95[1:42;16:26]
7:27[0:76;22:92]
9:93[0:94;39:05]
12:00[1:13;44:95]
4 51:44[40:22;61:80]
49:53[30:28;67:62]
3:94[0:88;11:42]
4:98[0:91;17:59]
8:18[1:11;34:22]
10:51[1:38;40:26]
8 59:47[48:63;69:49]
57:63[39:39;73:63]
8:49[2:39;18:76]
8:68[2:28;24;69]
11:70[2:62;40:93]
13:55[2:76;45:40]
20 68:47[59:29;76:19]
67:06[50:77;78:59]
19:76[10:88;30:39]
19:92[9:00;37:01]
22:62[8:71;51:64]
23:29[9:09;54:44]
40 71:08[62:36;77:72]
69:25[58:45;77:09]
35:37[23:82;45:38]
36:38[22:82;50:75]
39:35[22:30;62:07]
39:22[22:60;64:55]
80 70:70[53:55;80:23]
68:51[51:56;78:13]
52:57[39:69;62:13]
55:94[42:56;66:94]
58:99[43:80;74:47]
58:89[44:08;75:68]
200 67:37[35:78;86:03]
66:15[33:03;84:30]
71:31[53:73;79:70]
76:51[65:92;83:39]
79:39[69:94;87:60]
79:57[70:07;88:28]
Table 4: FEV share of output per capita due to technology shocks.
Max Share Long
Horizon HP1600 BP8;32 BP32;80 BP80;200 All run
1 2:34[0:29;7:18]
2:81[0:27;12:69]
20:67[9:52;35:22]
21:74[6:10;46:54]
22:59[3:32;57:14]
24:43[2:87;66:19]
4 13:14[5:40;22:92]
11:99[2:76;28:72]
7:72[2:33;18:41]
9:14[1:91;27:96]
11:97[2:06;40:86]
14:22[2:25;49:47]
8 20:58[10:08;32:31]
19:25[5:89;37:73]
4:73[1:94;13:19]
6:26[1:97;21:48]
9:75[2:31;35:38]
12:00[2:49;43:75]
20 22:70[11:32;34:95]
21:28[6:76;40:10]
4:22[1:57;11:89]
5:83[1:72;20:47]
9:27[2:23;34:35]
11:52[2:37;42:71]
40 22:99[11:42;35:37]
21:54[6:86;40:44]
4:12[1:46;11:83]
5:75[1:61;20:38]
9:23[2:18;34:28]
11:46[2:32;42:52]
80 23:09[11:42;35:45]
21:59[6:88;40:49]
4:11[1:43;11:84]
5:71[1:59;20:36]
9:25[2:17;34:27]
11:46[2:31;42:54]
200 23:10[11:42;35:50]
21:61[6:88;40:40]
4:11[1:42;11:86]
5:71[1:59;20:36]
9:24[2:17;34:23]
11:46[2:31;42:54]
Table 5: FEV share of hours per capita due to technology shocks.
16
1950 1960 1970 1980 1990 2000Â0.02
Â0.01
0
0.01
0.02
0.03
0.04Productivity growth
0 1 2 34
6
8
10
12x 10 Â5
1950 1960 1970 1980 1990 2000Â7.7
Â7.65
Â7.6
Â7.55
Â7.5
Â7.45
Â7.4Hours p.c.
0 1 2 30
0.005
0.01
0.015
0.02
Figure 1: Data: time series (�rst row) and sample spectral densities (second row).
17
10 20 30 40
Â0.5
0
0.5
1
1.5
Out
put p
.c.
Max Share  HP1600
10 20 30 40Â1
Â0.5
0
0.5
1
Hou
rs p
.c.
10 20 30 400
0.5
1
Prod
uctiv
ity
10 20 30 40
Â0.5
0
0.5
1
1.5
Max Share  BP8,32
10 20 30 40Â1
Â0.5
0
0.5
1
10 20 30 400
0.5
1
10 20 30 40
Â0.5
0
0.5
1
1.5
Max Share  BP32,80
10 20 30 40Â1
Â0.5
0
0.5
1
10 20 30 400
0.5
1
Figure 2: IRFs of hours, output, and productivity to various technology shocks (HP1600,
BP8;32, BP32;80).
18
10 20 30 40
Â0.5
0
0.5
1
1.5
Out
put p
.c.
Max Share  BP80,200
10 20 30 40Â1
Â0.5
0
0.5
1
Hou
rs p
.c.
10 20 30 400
0.5
1
Prod
uctiv
ity
10 20 30 40
Â0.5
0
0.5
1
1.5
Max Share  All freq.
10 20 30 40Â1
Â0.5
0
0.5
1
10 20 30 400
0.5
1
10 20 30 40
Â0.5
0
0.5
1
1.5
Long run
10 20 30 40Â1
Â0.5
0
0.5
1
10 20 30 400
0.5
1
Figure 3: IRFs of hours, output, and productivity to various technology shocks (BP80;200, all
frequencies, long run).
19
5 10 15 20 25 30 35 40Â0.5
0
0.5
1
1.5
2
Out
put p
.c.
Cholesky shock 1
5 10 15 20 25 30 35 40Â1
0
1
Hou
rs p
.c.
5 10 15 20 25 30 35 40
Â0.5
0
0.5
1
Prod
uctiv
ity
5 10 15 20 25 30 35 40Â0.5
0
0.5
1
1.5
2
Out
put p
.c.
Cholesky shock 2
5 10 15 20 25 30 35 40Â1
0
1
Hou
rs p
.c.
5 10 15 20 25 30 35 40
Â0.5
0
0.5
1
Prod
uctiv
ity
Figure 4: IRFs of hours, output, and productivity to Cholesky shocks.
20
Â3 Â2 Â1 0 1 2 30
1
2
3
4Max Share  HP
1600Max Share  BP
8,32Max Share  BP
32,80
Â3 Â2 Â1 0 1 2 30
0.5
1
1.5
2Max Share  BP
80,200Max Share  All freq.Long Run
Figure 5: Posterior distributions of the rotation angle between Cholesky shocks, �� =
atan2 (��2; ��1).
21